World food requirements will likely increase with dietary habits, such as higher meat and dairy consumption from rising per capita incomes. The pressure on global agriculture will increase, with demand for crops expected to double by 2050 roughly (Godfray et al., 2010; Tilman et al., 2011; United Nations, 2022). Even as we face these future burdens, there have been renewed investments in novel technologies to boost productivity. One of the most salient developments in global agriculture is the introduction of Genetically Modified (GM) crops (Aziz et al., 2022). GM crops, though, have been commercially adopted in both developed and developing countries over the past two decades, a fierce debate continues to rage concerning its implications for smallholder farmers in developing countries (Barrows et al., 2014; Abedullah et al., 2015; Bakhsh, 2017; Tokel et al., 2021; Evanega et al., 2022).

At the centre of this debate are the socioeconomic effects of Bacillus thuringiensis (Bt) cotton (Tabashnik et al., 2013; Zilberman et al., 2018; Smyth, 2020). The short-term benefits of Bt cotton are well documented, but the long-run impact is subject to substantial debate. Recent studies from China show positive benefits of Bt cotton in both the short-run and long-run, using farm-level and aggregate provincial-level data (Qiao, 2015; Qiao and Huang, 2020). Using farm-level data from India and Pakistan, studies have shown the short-term benefit (Subramanian and Qaim, 2009; Subramanian and Qaim, 2010; Subramanian et al., 2010; Bakhsh, 2017), but the long-term impact with aggregate data shows no gains for India (Kranthi and Stone, 2020). The debate about the sustainability of the Bt cotton gains is still a hot topic in India; thus, in this paper, I focus on India.

Even though farm-level studies have shown sizable gains in the first decade of Bt cotton adoption, these studies control for variables causing spurious associations affecting Bt adoption and yield growth (confounders) and selection bias from early Bt adopters being an unrepresentative group of progressive farmers. However, they suffer from cultivation bias due to preferential treatment for the costly Bt seeds in the initial adoption phase (Kranthi and Stone, 2020).

On the other hand, studies using aggregate data over two decades at the State- or National-level show mixed results (Plewis, 2019; Kranthi and Stone, 2020). These studies either do not distinguish late adopters from early adopters and non-adopters, thus suffering from selection bias or the analysis of aggregate data masks the effect size’s spatial heterogeneity. The benefits of Bt cotton can also change over time from pest pressure and resistance development, the effectiveness of sprays, availability of substitutes, and other dynamics (Kranthi and Stone, 2020; Lu et al., 2022). Yet, farm-level primary data studies addressing how individual farmers responded to the recent decade of Bt cotton adoption remain scarce.

From a policy perspective, it is interesting to examine the performance of Bt cotton, which has serious implications for the acceptance, regulatory approval, and adoption of GM food crops. For instance, the Chinese government has delayed commercialising many GM crops, including GM rice (Jin et al., 2019). Similarly, India has also put on hold the approval for three food crops, Bt mustard, potato and eggplant, pending further evidence on the impact of Bt cotton (Ramaswami and Pray, 2005; Jayaraman, 2010; Herring, 2015). Unfortunately, data nonavailability in India prevents a plot-level analysis of cross-state patterns that can address the above concerns.

This paper fills the gap using the newly collected detailed plot-level panel household data over five years in the recent decade from cotton farmers in the Ballari district of the Indian State of Karnataka. Because of the universal adoption of Bt cotton in the second decade of adoption, the data do not have non-Bt cotton plots. Though I cannot generalise the results, I dig deeper, applying trend and regression analysis to the performance of Bt cotton yields and profits. Since the farm-level data comes from the second decade of the Bt cotton adoption (Subramanian, 2018), I provide, unlike previous studies, new micro-level evidence controlling for confounders and addressing both selection bias and cultivation bias. Because of the near-universal adoption of Bt cotton in the study region in the second decade, we can dismiss the concerns arising from the differential adoption rate of Bt technology resulting in the selection bias.

To address the shortcomings of existing studies, I use comprehensive micro-level panel data collected from India’s second decade, 2012 to 2017, of Bt cotton adoption (Subramanian, 2018). The ethics committee at the Indian Institute of Management in Bangalore, India, approved the protocols related to the study. All methods were carried out following relevant guidelines and regulations. I obtained written informed consent from all the study participants. After obtaining ethical approval, I surveyed cotton-growing farmers in the Ballari district in Karnataka, a southwestern state of India. From the Bhoomi database, a census of land ownership in Karnataka, I randomly sampled 320 households. I followed a two-stage procedure. In the first stage, I identified all the villages predominantly growing cotton and randomly selected some households across these villages in the second stage.

I conducted four farm surveys among Indian cotton farmers between 2012 and 2017. A clustered random sampling procedure was followed to enlist the farmers from the Ballari district for the study. The first wave was implemented in March 2013, covering the 2012-2013 agricultural year. The second follow-up wave was implemented the following year, but the third and fourth waves were conducted consecutively after one gap year. The attrition is low except for the final year of the survey when some households disadopted cotton cultivation.

The dataset is an unbalanced panel of crop plots of varying plot sizes over four years. The samples selected are at the household level, and data collected is at the plot level though each sample household cultivated at least one cotton plot. However, a few households cultivated more than one plot. Thus, the number of cotton plots is higher than that of households. The estimation strategy is not at the household level but disaggregated by crop plots. The trained enumerators visited the sampled households at home and on the farm to administer the survey.

The farm survey, which includes a production module, collected retrospectively detailed plot level information on crop cultivation, such as outputs and inputs used. I collected detailed information on the number of family and hired labour used in the cost module, their days and hours worked, input quantity and prices, and transport costs. I recorded this information for each crop and every farming operation. There are 33 other crops grown that include cereals, pulses, and vegetable crops, including paddy, bengal gram, horse gram, maize, red gram, sugarcane, sunflower, cowpea, barley, groundnut, castor, green gram, and a combination of several crops raised together. Other information collected includes farmer-specific characteristics and household structures.

The attrition is low except for the final year of the survey when households disadopted cotton cultivation. Since our study has unbalanced panel data over four years, the fixed effects (FE) and random effects (RE) methods are used for estimating the impact of Bt cotton on crop yield and profits. I report the Hausman test to assess whether the FE model is the appropriate model for our data. As suggested by Wooldridge (2010) pooled ordinary least squares (OLS) method is employed when different samples are selected each year, thus not appropriate for our data. Though I control many confounding factors, the estimates are not causal, as are other studies in this debate. Addressing causality is empirically challenging when adoption is not wholly exogenous and evolves with the gains from the technology.

I estimate the following specification:

O i t is the outcome of interest (yield per acre, profit per acre) for household i in period t; S h i t is the share of Bt cotton area in total area cultivated by household i in time t; X i t is the plot and household level control variables, Y t is year fixed effects, δ v is group fixed effects, and ϵ i t is an error term. The effect of interest ( β 1 ) captures the average adoption impact of Bt technology on cotton yield and profits. A significant, positive value of β 1 indicates that yield and profits increase with the Bt cotton area.

The plot and household level controls include seed rate and the number of seeds in grams used per acre. The timing of sowing and harvest date in months, and the number of times the cotton plot was irrigated. I also include square terms as the control variables in the regressions to allow for nonlinear linkages between Bt cotton and input prices and quantity. I have pesticide quantity as a separate control variable to reflect infestation from sucking and chewing pests. Additional controls include age and education of the farmer in years, land owned, cultivated area and area under sharecropping in acres. The standard errors are clustered by household. I checked for heteroskedasticity in the data using scatter plots that do not show variations in outcome variables are more significant among large land size holding. Farmers who cultivated cotton had one or two equal size land plots. All the data analyses were conducted in Stata 17 (StataCorp).

The cotton crop is grown in the subtropical and seasonally dry tropical areas in the northern and southern hemispheres. According to the International Cotton Advisory Committee, the leading producing countries in 2021-22 are India (25%), China (25%), the United States (16%), Brazil (12%), and Pakistan (5%). Cotton has been an economically important commercial crop for India since the earliest times and grows all four cultivated cotton species (Blaise and Kranthi, 2020). In 2000, Gossypium hirsutum represented 69% of the total cotton in India, followed by G. arboreum (17%), G. herbaceum (11%), and G.barbadense (3%). India has pioneered the hybrid cotton technology and has become the only country where most of its acreage is under hybrids. The hybrid technology prevents seed saving and requires annual purchases of high-cost seed that leads to sub-optimal planting densities (Gutierrez, 2018). After the introduction of Bt hybrids for commercial cultivation in 2002-03, the composition of cultivation of species drastically changed. Presently, all the cotton in India is under the hirsutum group (>95%, 2012), leaving only less than 5% under arboretum and herbaceum (Government of India, 2017).

GM technology provides novel methods and capabilities to enhance agricultural productivity, mitigate its environmental footprint, and sustainably feed growing populations (Zilberman et al., 2018). Though Bt cotton is not a yield enhancing technology, it is designed to protect the yield potential of the variety that carries the trait from damage from some pests (Gutierrez et al., 2015). Yet, several controversies surround its impact, posing barriers to broader adoption and diffusion (Aziz et al., 2022). We can distinguish three sets of studies. (1) Short-term studies are primarily based on farm-level data. (2) Aggregate (provincial- or state-level) data showing long-term impact. (3) Long-run effect using farm-level data. Though the short-term studies show sizeable gains, these benefits can be offset by increased pesticide use and secondary pest outbreaks, thus raising doubts about the sustainability of the benefits of Bt cotton. Given the initial adoption phase of the technology, these studies suffer from selection and cultivation bias.

Most studies from Pakistan are based on cross-sectional single-year data suffering from self-selection and endogeneity issues. An exception to these studies is Bakhsh (2017) and Bakhsh et al. (2016), which use panel data for 2008 and 2009, though still in the initial phase of Bt cotton, show yields far less than previous studies. Though Bt cotton varieties in Pakistan were already under cultivation, it was only approved by the National Biosafety Committee in 2010, with further approvals in 2014 (Bakhsh et al., 2016).

Studies using aggregate data such as national and sub-national levels are divided on the economic gains from Bt cotton. Recent studies from China show increasing long-run economic benefits using both aggregate provincial-level data and farm-level panel survey data. However, studies based on aggregated secondary data from India over extended periods show modest benefits of Bt technology. Other studies show higher fertiliser use, better irrigation facility, and farmers’ bias towards the supply of other inputs to Bt relative to non-Bt cotton are the primary reason for the gains in cotton yields (Gutierrez et al., 2019).

Our results from the second decade using farm level panel data show yield sensitivity to pest pressure has increased, resulting in losses in some years and overall stagnation. Similar recent reports of pest attacks have been documented in other states, such as Gujarat, Madhya Pradesh, Maharashtra, Andhra Pradesh, and Telangana (Najork et al., 2021). Studies have demonstrated the significance of weather as an essential driver in heightening pest outbreak risks (Gutierrez et al., 2015; Zhang et al., 2018). These studies highlight changes in land use, climate and agricultural technologies that affect pest severity and management. The mirid bugs are likely to increase in severity with warmer temperatures and reduced insecticide spraying against bollworms. Climate variability is expected to underscore the challenges of meeting increasing global agricultural demand and sustainable development goals (Rosenzweig et al., 2001).

There is an urgent need to boost public investment in agriculture for GM technology to evolve in addressing the consequences of complex ecological dynamics between organisms and climate variability. However, some non-Bt cotton varieties offering superior results could also be part of the solutions provided to farmers (Gutierrez et al., 2020). For instance, studies have shown that adopting pure-line high-density short-season (non-Bt HD-SS) varieties of rainfed cotton could more than double current yields (Venugopalan et al., 2014; Kumar et al., 2020). It is also likely to avoid heavy pink bollworm infestations, thus reducing insecticides use. A more promising technology is the advanced molecular tools for precisely modifying plants using Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR/Cas9) (Hu and Li, 2022). This tool allows the plant breeders to make targeted sequence variations, resulting in rapid crop improvements (Aziz et al., 2022).

Previous microlevel studies across developing countries show sizeable gains from Bt cotton in the initial years of its adoption. Yet, recent studies from India using aggregate data show modest benefits over extended periods. In this paper, I use new farm-level panel data from the Indian district of Ballari to show yield sensitivity to pest pressure has increased in the second decade of adoption, resulting in losses in some years. More specifically, despite Bt technology that is claimed to be protecting from pink bollworms, farmers suffered massive yield losses from the pest. It represents a significant threat to the livelihoods and the very lives of millions of subsistence Indian cotton farmers.

Unlike recent evidence using both aggregate- and farm-level data from China showing that Bt cotton remains economically beneficial in the short and long run, our findings from India show economic benefits can diminish in the long run. It raises an important question on the sustainability of Bt cotton, even with the second-generation Bt gene. One of the reasons for the recurrence of the pink bollworm in India currently debated is the non-compliance of farmers with refuge requirements. With short-run profitability leading to increased adoption of Bt cotton and decreased natural refuge crops, the long-run outcome can be disastrous. Thus, policymakers might need to address non-compliance urgently and, in countries without refuge policies, rethink mandating a non-Bt cotton refuge.

Even though Pakistan and China currently do not have a refuge policy, China successfully reversed low levels of pink bollworm resistance by planting second-generation hybrid seeds from crosses between Bt and non-Bt cotton, naturally increasing the refuge area with non-Bt plants randomly interspersed within fields of Bt cotton (Tabashnik and Carrière, 2019). Though the efficacy of the built-in natural refuge with seed mixture appears to be successful in China, this strategy for managing pest resistance in other countries remains to be experimented with.

Though this study provides evidence from one district in Karnataka, I suggest establishing independent studies with representative surveys across the cotton-growing states to determine the extent of returns from Bt cotton in light of the widespread pink bollworm infestation. Since Bt cotton is the only GM crop technology widely adopted by smallholder farmers, the findings can contribute to the broader public debate on the future of agricultural biotechnology in developing countries. This paper can inform the future scientific development of GM technology, which is expected to address food insecurity in the face of climate change.

The aggregate data used for the study and the STATA codes for the statistical analysis based on the regression model is freely available online at https://reshare.ukdataservice.ac.uk/853079/.

The author confirms being the sole contributor of this work and has approved it for publication.